Valdivia Tech
It's been a long time since I wrote — I apologize, and I apologize to myself. I've been working on something that not long ago was just a concept. Today it's a company with a site: valdivia.tech. This post is about what's inside, because I've spent months building things that until now only existed as scattered entries on this blog.
The thesis
Valdivia Tech is an AI agent lab for electric power systems. My bet is simple: the power sector carries an enormous amount of heavy lifting — repetitive studies, regulatory paperwork, document review, infrastructure operation, and so on — that eats the time of engineers who should be doing the important work. Our agents take the heavy lifting. An engineer signs the final decision.
I already wrote about the MIT study that found 95% of enterprise AI pilots produce no measurable impact. Models keep getting better and the results still don't show up. My conclusion after months of work is that the gap doesn't close with a bigger model — it closes working directly inside the operation and, why not, with a better harness: understanding how a request gets classified, what gets reviewed, where it hurts. That's why the lab runs in FDE mode (I'll write about Forward Deployed Engineers in an upcoming post): today I'm building agents for clients in different regions of the country, sitting inside their processes.
What has already left the lab
Right now I have three agents in production, and each one has a public history on this blog.
Don Nelson runs electrical studies end to end: it takes the request, runs PowerFactory, and returns the report within 72 business hours. It's where I started — I set myself the challenge of getting there in 60 days, then extended it to 100. I'll go deeper into Don Nelson's current state and harness in an upcoming post. But it is OFFICIALLY doing studies submitted to the CEN 👀.
EDITH is the sovereign AI layer: it operates on-prem inference infrastructure — it watches the GPUs and acts on the hardware where the AI runs. It was born as the computer I built for electrical simulations and ended up with a nervous system of its own.
Mateo is the grid model served over an API: it reconstructs the substations of the Chilean national grid and can answer where every data point came from. It's the least flashy piece and the one the other two use the most.
Applied research
The part of the site I'm proudest of is the research section, because a lab that doesn't publish is a consultancy with a different name. And this is key: research is something Valdivia Tech will keep doing.
The first piece is the Spark paper: a harness where the agent writes and executes Python scripts that drive a power-system simulator, and where procedural knowledge lives in verifiable recipes distilled from verified frontier-model runs — not in the model's weights.
The result: a free, local 26B model on a single consumer GPU collapses to 32% task success without the harness and reaches 100% with it — indistinguishable from the frontier model.
To me, that paper is the whole thesis in miniature: the value isn't in the model — it's in the harness, the recipes, and the domain knowledge. Which is exactly what a lab embedded in the operation can build and a generic vendor cannot.
What's next
I'm going to start publishing on my blog again. It filled my heart and I did it because I loved it. Maybe building in public got more complicated: I started touching topics I couldn't write about, because they reached places that simply couldn't be published.
If you work in the power sector and heavy lifting is eating your team, let's talk: valdivia.tech. We'll tell you if there's an agent that can take it. Or if there isn't one yet.